Thread per-WSG cw/mad habitat model through classify (#286)

[NewGraphEnvironment]

Summary

• Every bundle declares a per-watershed-group habitat model table, parameters_habitat_method.csv (cw or mad). lnk_pipeline_classify() passes it to fresh::frs_habitat_classify(params_method =), and a new method_csv = argument overrides it. A custom bundle without the file falls back to fresh's copy, which is hashed into config_hash.
• Shipped tables are all cw: a frozen copy of smnorris/bcfishpass@1fae4ea parameters/example_newgraph. They are deliberately not csv-synced. sync_bcfishpass_csvs.R selects on source:, and byte drift auto-merges, so an upstream cw→mad flip would otherwise have changed default unreviewed.
• mad_m3s is joined onto the working streams from fwa_stream_networks_discharge. It is never persisted, so the persist shape is unchanged. log_input now fingerprints the discharge table.
• A mad group skips the stream-order rearing bypass; bcfishpass applies it only in its cw branch.
• fresh is pinned at v0.36.2, with the floor at 0.35.0. lnk_preflight_fresh() gains required_formals and asserts the params_method argument rather than a version.
• fresh 0.36.0 reversed frs_db_conn()'s env precedence. data-raw/wsg_vignette_data.R therefore moves to lnk_db_conn().

Related Issues

• Closes Thread fresh params_method (per-WSG cw/mad) through lnk_pipeline_classify #286
• Relates to NewGraphEnvironment/sred#24

Test plan

• devtools::test(): FAIL 0, PASS 2298, WARN 16 (the warning baseline).
• devtools::check(): 0 errors. The 3 warnings and 3 notes were already there; no added line is non-ASCII.
• Live no-change proof (local fwapg, scratch schemas, config default). The per-species streams_habitat digests are identical across branch, main + fresh 0.36.2, and main + fresh 0.33.0, on ADMS (5 species) and BULK (7 species). ADMS also matches with mad_m3s dropped.
• ADMS on mad. No stream segment falls outside its MAD range, and BT/RB lose all stream habitat. BULK on mad (no discharge coverage) loses all stream habitat, without an error.
• Mutation checks. A stubbed formals check turns 3 preflight tests red, and dropping the fallback hash turns test-lnk_log.R red.
• Code review: 3 /code-check rounds plus a plan review. One real bug was found (the csv-sync enrollment) and fixed.

Evidence: data-raw/logs/params_method_286/. Mechanics: RUNBOOK §7, "Channel width or discharge, per watershed group".

Notes

• Re-prep cyphers before the next dispatch. An image on fresh < 0.35 fails preflight, by design.
• config_hash changes for every bundle, because a new file is declared. No model output changed.
• Not done here:
  • lnk_habitat_validate() is still cw-only. A follow-up issue body is drafted in planning/archive/2026-10-issue-286-params-method/findings.md and has not been filed.
  • fresh's own bundled method table is missing PINE (187 vs 188 rows). That has no effect, since an unlisted group is cw.

Draft NEWS (for the release commit):

Each config can now choose channel width or discharge per watershed group (#286). Bundles declare parameters_habitat_method.csv, which lnk_pipeline_classify() passes to fresh as params_method; every shipped bundle is all cw, so outputs are unchanged. A mad group classifies on mad_m3s, joined onto the working streams (never persisted), and skips the stream-order rearing bypass. Requires fresh >= 0.35.0 (pinned at v0.36.2), and lnk_preflight_fresh() now checks for the argument, so re-prep cypher images before the next run. config_hash changes for every bundle.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WJE2szSCkeV9R3qKhuSwyx